Return to search

Dynamic allocation of resources using machine learning and quantile regression by harnessing the power of software defined networks

In the last decade, data networks have shifted from the static deployment of resources to a dynamic approach. With the help of Software Defined Networks (SDN) and Network Function Visualization (NFV), information and data about the network can be collected. Also, deployment and allocation of resources can be delegated to a central controller. In this thesis we investigate the power of SDN and how central management of resources can help produce better and efficient data networks. It begins with an introduction to SDN and its capabilities. The added benefits of SDN over traditional network frameworks and topics that SDN contributed most to. We show the power of collecting data using SDN and how it enables different approaches to accomplish the needed task. This was facilitated by the programmability and the separation of the control and data planes. We tackle the simple task of measuring the delay between two communicating devices in the network. The results show that SDN is capable of providing a rich infrastructure to build future networks. Also, it illustrates that using SDN to measure the delay between devices in the network can give accurate results. The differences between the tested techniques is shown and evaluated. After collecting the data from the network, the next step is getting an insight on that data. Next we used collected network bandwidth data to predict future bandwidth usage. We used various prediction models to establish prediction intervals. We created a state of the art metric that evaluates and compares the performance of each model. We show that the network bandwidth is highly predictable and that dynamic allocation of network bandwidth is attainable. The next logical step is to act upon those insight which is investigated next. We establish the same prediction models investigated but instead of prediction intervals we establish upper quantiles. Prediction is done on data center resources data set. The results show that using quantile prediction can give guarantees on resources usage boundaries which implies a guarantee on service level agreements. Allocating just the needed resources, produce a more efficient data center and in turn cuts a lot of the needed energy. Our estimate show that upto 56% of power can be saved without violating the service level agreement. / Graduate

Identiferoai:union.ndltd.org:uvic.ca/oai:dspace.library.uvic.ca:1828/13927
Date02 May 2022
CreatorsAlutaibi, Ahmed
ContributorsGanti, Sudhakar
Source SetsUniversity of Victoria
LanguageEnglish, English
Detected LanguageEnglish
TypeThesis
Formatapplication/pdf
RightsAvailable to the World Wide Web

Page generated in 0.0022 seconds